Description

Machine Learning HW5
Outline
1. Machine translation
2. Workflow
3. Training tips
4. Requiements
5. Report
6. JudgeBoi Guide
7. Regulation and Grading policy
Machine Translation
Machine Translation
In this homework, we’ll translate English to Traditional Chinese e.g.
● Thank you so much, Chris. -> 非常謝謝你，克里斯。
Since sentences are with different length in different languages, the seq2seq framework is applied to this task.

Training datasets
● Paired data
○ TED2020: TED talks with transcripts translated by a global community of volunteers to more than 100 language
○ We will use (en, zh-tw) aligned pairs
● Monolingual data
○ More TED talks in traditional Chinese
source: Cats are so cute Evaluation target: 貓咪真可愛
output: 貓好可愛
BLEU
● Modified n-gram precision (n = 1~4)
● Brevity penalty: penalizes short hypotheses
c is the hypothesis length, r is the rerferencwe length
● The BLEU socre is the geometric mean of n-gram precision, miltiplied by brevity penalty
Workflow
Workflow

Workflow
1. Preprocessing
a. download raw data
b. clean and normalize
c. remove bad data (too long/short)
d. tokenization 2. Training
a. initialize a model
b. train it with training data
3. Testing
a. generate translation of test data
b. evaluate the performance
Training tips
Training tips
● Tokenize data with sub-word units
● Label smoothing regularization
● Learning rate scheduling
● Back-translation

Tokenize
● Tokenize data with subword units
○ Reduce the vocabulary size
○ Alleviate the open vocabulary problem
○ example
■ ▁put ▁your s el ve s ▁in ▁my ▁po s ition ▁.
■ Put yourselves in my position.
Label smoothing
● Label smoothing regularization
○ When calculating loss, reserve some probability for incorrect labels ○ Avoids overfitting

Leaning rate scheduling
● Learning rate scheduling
○ Increasing the learning rate linearly for the first warmup_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number.
○ Stabilizing training for transformers in early stages

Back translation
Using monolingual data for creating synthetic translation data
1. Train a translation system in the opposite direction
2. Collect monolingual data in target side and apply machine traslation
3. Use translated and original monolingual data as additional parallel data to train stronger translation systems
Back translation
Some points to note about back-translation
1. Monolingual data should be in the same domain as the parallel corpus
2. The performance of the backward model is critical
3. Increase model capacity since the data amount is increased

Requirements
Baselines
Baseline Public score Estimated time(kaggle)
Simple 14.58 1 hour
Medium 18.04 1 hour 40 mins
Strong 25.20 ~3 hours
Boss 29.13 > 12hours
Baseline Guide
● Simple Baseline: Train a simple RNN seq2seq to acheive translation
● Medium Baseline: Add learning rate scheduler and train longer
● Strong Baseline: Switch to Transformer and tuning hyperparameter
● Boss Baseline: Apply back-translation
Simple Baseline
Train a simple RNN seq2seq to acheive translation
● Running the sample code should pass the baseline
Medium Baseline
Add learning rate scheduler and train longer

config = Namespace(
. . .
# maximum epochs for training max_epoch=15, # medium: -> 30 start_epoch=1,
. .
.
)

def get_rate(d_model, step_num, warmup_step):
# TODO: Change lr from constant to the equation shown above lr = 0.001 return lr

Strong Baseline
Switch to Transformer and tuning hyperparameter
encoder = RNNEncoder(args, src_dict, encoder_embed_tokens) decoder = RNNDecoder(args, tgt_dict, decoder_embed_tokens)
-> # encoder = TransformerEncoder(args, src_dict, encoder_embed_tokens)
# decoder = TransformerDecoder(args, tgt_dict, decoder_embed_tokens)
arch_args = Namespace( encoder_embed_dim=256, encoder_ffn_embed_dim=512,
encoder_layers=1, # recommend to increase -> 4 decoder_embed_dim=256, decoder_ffn_embed_dim=1024,
decoder_layers=1, # recommend to increase -> 4 share_decoder_input_output_embed=True, dropout=0.3, for othertransformer-base hyperparameters for, pleaserefer to
) Table 3 in Attention is all you need
Boss Baseline
Apply back-translation
1. Train a backward model by switching languages
a.source_langasd = “zh”,
b.target_langasd = “en”,
2. Translate monolingual data with backward model to obtain synthetic data
a. Complete TODOs in the sample code
b. All the TODOs can be completed by using commands from earlier cells
3. Train a stronger forward model with the new data
a. If done correctly, ~30 epochs on new data should pass the baseline
Links
Colab sample code
Kaggle sample code
JudgeBoi

Report
Report Overview
● Problem 1
○ Visualize the similarity between different pairs of positional embedding and briefly explain the result.
● Problem 2
○ Clip gradient norm and visualize the changes of gradient norm in different steps. Circle two places with gradient explosion.
Problem 1: Visualize Positional Embedding
Given a (N x D) positional embeding lookup table, you aim to get a (N x N) “similarity matrix” by calculating similarity between different pairs of embeddings in the table.
You need to visualize the similarity matrix and briefly explain the result.
In this problem, we focus on the positional embeddings of the decoder
Problem 1: Similarity Matrix
p1 p2 p3 p4 p5
In the sence of encoding
1 0.8 0.6 0.4 0.3
0.8 1 0.8 0.6 0.4
0.6 0.8 1 0.8 0.6
0.4 0.6 0.8 1 0.8
0.3 0.4 0.6 0.8 1
p1positional information, we
expect that the similarity
p2between the embedding of
close positions is stronger.
p3 p4 p5
Problem 1: Cosine Similarity
We recommend you to measure the similarity between two vectors by cosine similarity.
There is a pytorch implementation of cosine similarity. Check more detail in the following link.
https://pytorch.org/docs/stable/generated/torch.nn.functional.cosine_similarity.html

Problem 1: Tips and Hint
You could get the positional embeddings of decoder by following codes
pos_emb = model.decoder.embed_positions.weights.cpu().detach()
torch.Size( [1026, 256] )
Problem 2: Gradient Explosion

Problem 2: Clipping Gradient Norm
1. Set up a maximum norm value max_norm
2. Collecting the gradient of each parameters to be a vector. Calculate the p-norm of the vector to be Lnorm
3. If Lnorm <= max_norm, do nothing. Otherwise calculate the scale factor scale_factor = max_norm / Lnorm and multiply each gradient by the scale factor.

p-norm 2-norm
Problem 2: Visualize Gradient Norm
Step1: Apply clips gradient norm and set max_norm = 1.0 .
Step2: Make a plot of “gradient norm v.s step”.
grad norm v.s. step

Problem 2: Visualize Gradient Norm
Step3: Circle two places with gradient explosion (where the clip_grad_norm function take effect)
grad norm v.s. step
Pytorch: clip_grad_norm_()
Problem 2: Overview
In this problem, you need to do
1. Plot the grad_norm
defa. train_one_epochasd (epoch_itr, model, task, criterion, optimizer, accum_steps=1):
b. asd .
c. asd .
d. asd .
e. asd optimizer.multiply_grads(1 / (sample_size or 1.0)) # (sample_size or 1.0) handles
the case of a zero gradientf. asd
g. asdgnorm = nn.utils.clip_grad_norm_(model.parameters(), config.clip_norm) # grad norm clipping prevents gradient exploding -> this is the grad_norm for every step
h. asd
2. Circle two place with gradient explosion (if there is gradient explosion)

JudgeBoi Guide
Previously… Github Account Survey
We have kindly requested everyone to report your github username and ID.
IMPORTANT: You must take this survey in order to submit to JudgeBoi server.
Step 1: Register for Submission

Step 2: Sign-in with Github
● If you have not completed the Github account survey
○ You can log in
○ You will not be able to submit
fill in username > fill in password >
Step 3: Submit your Results
You can now submit results to the server and view the leaderboard.
1) Click
3) Check the leaderboard
2) Submit results
Step 4: Select your submissions
● You can select up to 2 submissions.
● If none of your submissions is chosen, we will use the submission with the best public score.
● If your selection is successful, you will see a message box as follows:

JudgeBoi Rules
● 5 submission quota per day, reset at midnight.
○ Users not in the whitelist will have no quota.
● Only *.txt file is allowed, filesize should be smaller than 700kB.
● The countdown timer on the homepage is for reference only.
● We do limit the number of connections and request rate for each IP.
○ If you cannot access the website temporarily, please wait a moment.
○ If this prevents uploads, we do not offer additional opportunities for remediation
● Please do not attempt to attack JudgeBoi.
● For any JudgeBoi issues, please post on NTUCOOL discussion
○ Discussion Link: https://cool.ntu.edu.tw/courses/11666/discussion_topics/91777

Regulations and Grading Policy
Grading
● simple (public) +0.5 pts
● simple (private) +0.5 pts
● medium (public) +0.5 pts
● medium (private) +0.5 pts
● strong (public) +0.5 pts
● strong (private) +0.5 pts
● boss (public) +0.5 pts
● boss (private) +0.5 pts
● code submission +2 pts
● report +4 pts
Total : 10 pts
Code Submission
● NTU COOL (4pts)
○ Compress your code and report into
<student ID>_hwX.zip
* e.g. b06901020_hw1.zip * X is the homework number
○ We can only see your last submission.
○ Do not submit your model or dataset.
○ If your code is not reasonable, your semester grade x 0.9.
Code Submission
● Your .zip file should include only
○ Code: either .py or .ipynb
○ Report: .pdf (only for those who got 10 points)
● Example:

Report Submission
Answer the questions on GradeScope
Deadlines
Regulation
● You should NOT plagiarize, if you use any other resource, you should cite it in the reference. (＊)
● You should NOT modify your prediction files manually.
● Do NOT share codes or prediction files with any living creatures.
● Do NOT use any approaches to submit your results more than 5 times a day.
● Do NOT search or use additional data or pre-trained models.
● Your final grade x 0.9 if you violate any of the above rules.
● Prof. Lee & TAs preserve the rights to change the rules & grades.
If any questions, you can ask us via…
● NTU COOL (recommended)
○ https://cool.ntu.edu.tw/courses/4793
● Email
○ The title should begin with “[hwX]” (X is the homework number)